Roboticist here. All of this, and he didn’t mention compliance or online adaptation to otherwise un-sensable dynamics. Or massively complex miniature mechanisms.
Current generation tactile sensors cost a couple thousand $ PER FINGER, and have a real world MTBF of hours. The cost can be solved with economy of scale. The fragility is harder.
UltraSane 36 minutes ago [-]
I'm picturing humanoid robots having to operate in pairs so they can constantly fix each other.
kooi 30 minutes ago [-]
Yes the problem is very hard. Mainly because high DOF generalization is very difficult.
We have self driving cars because what are the control inputs? Pedal, brake, steering wheel. This already took many many years.
Now for a humanoid robot: An action space that is metaphorically Hilbert. (Physically, yes, obviously)
Also, IMO, LLM's can aid the development of robots, but do little beyond a planning, human control interface. Below that it's the domain of control and the solution will be the correct combination of classical, neural, and real time optimization based control.
All the bad-ass biped robots that actually look natural? It's PID controls wrapped with control barrier functions constraining the QPs that are being solved in real time.
But that's annoying to derive per-application. So we'll need neural methods which can be learned (while being constrained by a priori knowledge of dynamics). My hunch is that the Yann LeCunn type of jepa models will be how tasks can be learned.
UltraSane 37 minutes ago [-]
The physical world is basically an infinite amount of global state that must be perceived indirectly through imperfect sensors and acted on using imperfect motors and manipulators.
logicallee 2 hours ago [-]
people very soon as going to look back at all of us and just think "they didn't even have robots yet! how did they even eat?" Well, future readers, it isn't easy! We either go to the store and bring back groceries ourselves or pay a helper to do it. No robot in the loop there. Then when it comes time to cook something we either have to cut it up and otherwise prepare it and then cook it ourselves, or another human can do it, or we can buy fully prepared foods that don't require cooking. And then when we're done we have to put our own dishes away into a dishwasher. We can't just say a few words and have a robot do it for us!
Rendered at 01:25:51 GMT+0000 (Coordinated Universal Time) with Vercel.
Current generation tactile sensors cost a couple thousand $ PER FINGER, and have a real world MTBF of hours. The cost can be solved with economy of scale. The fragility is harder.
We have self driving cars because what are the control inputs? Pedal, brake, steering wheel. This already took many many years.
Now for a humanoid robot: An action space that is metaphorically Hilbert. (Physically, yes, obviously)
Also, IMO, LLM's can aid the development of robots, but do little beyond a planning, human control interface. Below that it's the domain of control and the solution will be the correct combination of classical, neural, and real time optimization based control.
All the bad-ass biped robots that actually look natural? It's PID controls wrapped with control barrier functions constraining the QPs that are being solved in real time.
But that's annoying to derive per-application. So we'll need neural methods which can be learned (while being constrained by a priori knowledge of dynamics). My hunch is that the Yann LeCunn type of jepa models will be how tasks can be learned.